Context. For cloud-native applications, the choice of scaling method can highly impact both the application performance and operational cloud costs. To be able to evaluate different cloud application architectures at design time, an architecture can be modeled and simulated. Palladio [BKR07] is a modeling language and simulator that allows the modeling of architectures for performance predictions at design time. The Slingshot approach [KKSB23] proposes adding a new type of architectural view called Scaling Policy Definitions (SPDs) to Palladio for model-based performance predictions of elasticity policies. In addition, the Slingshot approach encompasses the Slingshot simulator which employs discrete event simulations of fixed length to determine the quality impact of modeled policies. Problem. Resource provisioning and autoscaling in the cloud may rely on advanced scaling approaches that are not only reactive but also include predictive components that are often learningbased. However, the Slingshot approach currently supports only the modeling of elasticity policies that can be classified as reactive. Objective. The objective of this thesis is to identify currently used learning-based elasticity policies and to implement them in Slingshot. Thus, this thesis extends the Slingshot approach with support for learning-based approaches and compares them both to each other and their Slingshot implementation with results provided in the original paper. Method. The approaches are identified using a rapid literature review. For the extension of SPDs, a model-driven software development approach is followed. The evaluation is based on the evaluations done by the identified papers and is performed using Slingshot. Result. The SPD metamodel is extended with support for learning-based elasticity policies and the Fuzzy SARSA and Fuzzy Q-Learning approaches are implemented in the Slingshot simulator. The evaluation shows that these policies perform better than the existing reactive scaling policies. Conclusion. By implementing state-of-the-art predictive elasticity policies in Slingshot and extending SPDs with support for modeling them, this thesis makes Slingshot and the SPD language more complete. The evaluation shows that learning-based elasticity policies can offer improved adherence to Service Level Objectives (SLOs) while using fewer resources, but initially perform worse than reactive policies during the learning phase.